Getting forked by Microsoft

The author of the open-source project Spegel discovered that Microsoft had created a similar project called Peerd, which appeared to be a fork of Spegel, with Microsoft acknowledging Spegel as an inspiration but not contributing changes back to the original project. The author feels that Microsoft's actions have created confusion among users and made it difficult for Spegel to gain recognition, leading to questions about how sole maintainers can work with large corporations without being taken advantage of.

Read more

LLM-powered tools amplify developer capabilities rather than replacing them

The author used Claude Code, an AI-powered coding tool, to build two apps and generated approximately 30,000 lines of code, learning that such tools are not replacements for human programmers, but rather amplifiers of their abilities, similar to a mech suit. The author found that while AI can greatly accelerate certain aspects of development, it requires constant vigilance and human oversight to maintain architectural integrity and make key decisions, and that experience and wisdom are still essential in determining when to discard and rebuild code.

Read more

AI assisted search-based research works now

AI-assisted search tools have improved significantly in 2025, with models like o3 and o4-mini from OpenAI and Gemini 2.5 Pro from Google demonstrating the ability to accurately search and provide reliable information without hallucinating details. These tools can now run searches as part of their reasoning process, allowing for more accurate and trustworthy results, and have the potential to revolutionize tasks such as research and code porting.

Read more

Local LLM inference – impressive but too hard to work with

Local LLM inference has made tremendous progress, allowing powerful language models to run in browser tabs for free, but it's not yet ready for production as a developer platform due to limitations in finding and deploying the correct model for a given task. Despite this, local inference offers benefits such as cost savings, improved privacy, speed, and enabling offline use, and frameworks like llama.cpp, Ollama, and WebLLM are working to make local LLM inference more accessible and efficient.

Read more

A weird phrase is plaguing scientific papers

The term "vegetative electron microscopy" is a nonsensical phrase that originated from a combination of errors in digitizing and translating scientific papers, and has since been perpetuated and amplified by artificial intelligence (AI) systems. This "digital fossil" has become a permanent fixture in our knowledge ecosystem, appearing in 22 papers and highlighting the challenges of identifying and correcting errors in AI systems, which can have significant implications for the integrity of scientific research and publishing.

Read more

Learnable Multi-Scale Wavelet Transformer: A Novel Alternative to Self-Attention

The Learnable Multi-Scale Wavelet Transformer (LMWT) is a novel architecture that replaces traditional self-attention with a learnable multi-scale Haar wavelet transform module, allowing it to efficiently capture local and global context in sequences. The LMWT achieves competitive performance with substantial computational advantages, making it a promising alternative for efficient sequence modeling, as demonstrated through experiments on a machine translation benchmark.

ArXiv

Assessing Computer Science Student Attitudes Towards AI Ethics and Policy

The attitudes and competencies of post-secondary computer science students towards AI ethics and policy are of particular interest, as they will play key roles in the development and deployment of future AI innovations, yet their attitudes remain understudied. A survey and follow-up interviews with CS students at a US university were conducted to assess their attitudes towards AI ethics and policy, with the findings drawing parallels and contrasts to broader public opinion polling and implications for the future of AI education and governance.

ArXiv

In between myth and reality: AI for math – a case study in category theory

Researchers have been conducting various tests to evaluate the performance of AI systems in solving math problems, yielding mixed results. This paper discusses an experiment involving two prominent AI systems, aiming to understand how AI can assist mathematical research and provide suggestions for improvement to AI system developers.

ArXiv

Machine learning with neural networks (2021)

The lecture notes cover a machine learning course on neural networks, divided into three parts: stochastic recurrent networks, supervised learning, and learning from unlabeled data sets. The course aims to explain the fundamental principles of neural networks, covering topics such as Hopfield networks, multilayer perceptrons, convolutional neural networks, unsupervised learning, and reinforcement learning, with a focus on common concepts and ideas throughout.

ArXiv

Ultra-precision formation flying demonstration for space-based interferometry

SILVIA is a proposed mission concept that aims to demonstrate ultra-precision formation flying between three spacecraft, achieving sub-micrometer precision in relative distance control through advanced sensors and propulsion systems. The mission will help bridge the technology gap for future space-based observatories, such as gravitational wave detectors and exoplanet imagers, by demonstrating the necessary precision and control in a low Earth orbit setting.

ArXiv

Show HN: Light like the Terminal – Meet GTK LLM Chat Front End

GTK LLM Chat is a graphical interface for interacting with Large Language Models (LLMs) built with GTK, offering features such as multiple conversation support, real-time streaming responses, and markdown rendering. The application can be installed using pipx and llm, and its usage involves running the application with optional arguments to specify conversation IDs, system prompts, and model selections.

Repo

Show HN: HyperAgent: open-source Browser Automation with LLMs

Hyperagent is a browser automation tool that combines Playwright with AI capabilities, allowing for intelligent automation of browser tasks using large language models (LLMs). It provides features such as AI commands, stealth mode, and fallback to regular Playwright, and can be customized with different LLM providers, output schema definitions, and custom actions.

Repo

GPT-crawler – generate knowledge files to create your own custom GPT from a URL

The GPT Crawler is a tool that allows users to crawl a website and generate knowledge files to create a custom GPT model from one or multiple URLs. By configuring the crawler and running it, users can generate a file called output.json that can be uploaded to OpenAI to create a custom assistant or custom GPT, providing UI or API access to the generated knowledge.

Repo

Fixing the Biggest Flaw in GPT-4o Image Gen

FaceEnhance is a tool for improving facial consistency and quality in AI-generated images, allowing users to dramatically enhance facial fidelity while preserving the original image's background, lighting, and composition. The tool can be installed and run using Python, with a simple web interface available through Gradio, and requires a GPU with at least 48GB VRAM and 60GB of free disk space to function.

Repo

Show HN: I Open-Sourced My AI Toy–ESP32, OpenAI Realtime API on Deno, Supabase

ElatoAI is a real-time AI speech project that enables uninterrupted global conversations for over 10 minutes, powered by OpenAI's Realtime API, ESP32, Secure WebSockets, and Deno Edge Functions. The project consists of three main components: a Next.js frontend, Deno Edge Functions, and an ESP32 Arduino client, allowing users to create custom AI agents and engage in conversations with them.

Repo